In the early stages of the generative AI boom, the volume of 'tokens'—the basic units of text processed by large language models—was frequently used as a shorthand for a company’s innovative vigor. For high-growth tech firms, burning through massive token counts was seen as a sign of aggressive adoption and technological leadership. However, that consensus is rapidly unraveling as corporate giants realize that high consumption does not necessarily translate into high productivity.
Amazon recently signaled a sharp pivot by shutting down 'Kirorank,' an internal leaderboard used by developers to track their token usage. Originally intended to encourage the integration of AI into business workflows, the system backfired by incentivizing 'token maxxing.' Employees, eager to climb performance rankings, began deploying AI agents in redundant or low-value tasks to inflate their metrics, leading to a massive waste of expensive computational resources. Dave Treadwell, Amazon’s Senior Vice President, cautioned staff that the era of using AI for its own sake is over.
Amazon is not alone in its retreat from vanity metrics. Meta recently shuttered a similar internal tracker nicknamed 'Claudeonomics' after data revealed that employees had consumed over 60 trillion tokens in a single month—a volume that would cost nearly $900 million at market rates. Insiders reported that the output from the top 'token burners' was largely 'disposable garbage,' offering little to no long-term value to the company’s core objectives. This disconnect highlights a growing anxiety among tech executives: the inability to find a clear correlation between AI spending and financial returns.
Uber has also hit a wall, with Chief Operating Officer Andrew Macdonald noting that the ride-hailing giant has yet to see productivity gains that justify its surging AI investments. Earlier this year, Uber’s technical leadership admitted to exhausting its annual budget for Anthropic’s Claude Code by April, a stark reminder of how quickly costs can spiral without disciplined oversight. Even OpenAI CEO Sam Altman has acknowledged the growing skepticism, noting that while companies feel more 'productive,' they are increasingly asking where the actual revenue growth is hiding.
Industry analysts are now urging a shift toward outcome-oriented metrics, such as the cost per successful code merge, rather than raw consumption. The emerging consensus suggests that the most efficient path forward involves moving the majority of the workforce into a 'medium-use' tier. By avoiding both the stagnation of AI under-utilization and the fiscal recklessness of token maxxing, firms hope to finally bridge the gap between AI experimentation and genuine economic value.